Reproduce “Scaling Laws For Dense Retrieval” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yan Fang, Jingtao Zhan, Qingyao Ai, Jiaxin Mao, Weihang Su, Jia Chen, Yiqun Liu
This paper establishes that the performance of dense retrieval models follows power-law scaling relationships with respect to model size and training data volume. It introduces contrastive entropy as a continuous evaluation metric and demonstrates how these scaling laws can guide resource allocation for efficient system development.
Key insights
- Dense retrieval model performance adheres to precise power-law scaling relationships with both model size ($L(N) = A/N^\alpha + \delta_N$) and annotated data size ($L(D) = B/D^\beta + \delta_D$).
- An irreducible loss term ($\delta$) is consistently observed, indicating inherent limitations in ultimate performance due to factors like data noise or task subjectivity.
- Annotation quality affects the scaling efficiency, with LLM-generated data showing potential to surpass human annotations at sufficient scales for certain tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.